- Messages
- 33,139
- Country
To test your data I would need at least one bigger image to test on. If you have a few dozen samples already that would be great, but I can extract these myself as well. In general you won't need thousands of samples.
-
Which the release of FS2020 we see an explosition of activity on the forun and of course we are very happy to see this. But having all questions about FS2020 in one forum becomes a bit messy. So therefore we would like to ask you all to use the following guidelines when posting your questions:
- Tag FS2020 specific questions with the MSFS2020 tag.
- Questions about making 3D assets can be posted in the 3D asset design forum. Either post them in the subforum of the modelling tool you use or in the general forum if they are general.
- Questions about aircraft design can be posted in the Aircraft design forum
- Questions about airport design can be posted in the FS2020 airport design forum. Once airport development tools have been updated for FS2020 you can post tool speciifc questions in the subforums of those tools as well of course.
- Questions about terrain design can be posted in the FS2020 terrain design forum.
- Questions about SimConnect can be posted in the SimConnect forum.
Any other question that is not specific to an aspect of development or tool can be posted in the General chat forum.
By following these guidelines we make sure that the forums remain easy to read for everybody and also that the right people can find your post to answer it.
P3D v5 Water Detection idea (work-around)
- Thread starter Clutch Cargo
- Start date
- Messages
- 1,588
- Country
Ok, I will prep a few images for you to download.
- Messages
- 1,588
- Country
Arno, I have uploaded two images for you to download and test with. Each one is 1GB+ in size. Email I sent should have the link. One image is in Redding, California and the 2nd is in the state of Washington. Both I think will make good test areas as the water has various colors, rapids, etc. Let me know how it goes. 
- Messages
- 33,139
- Country
Hi,
I ran a quick test of the CA image with the new Artificial Neural Networks - Multi-Layer Perceptrons (MLP) step and I must say that I am quite pleased with the first results:
There are still some issues in areas where the water is much darker and with the white dots in the water (waves I guess). Below are some zoomed in parts that show these trouble areas.
I did have to use the MRS instead of KMean clustering for the detection of the segments. I know you prefer KMeans due to performance, but that method always combined water and shadow in one segment near one of the dams. In that case they will also be detected as the same.
I ran a quick test of the CA image with the new Artificial Neural Networks - Multi-Layer Perceptrons (MLP) step and I must say that I am quite pleased with the first results:
There are still some issues in areas where the water is much darker and with the white dots in the water (waves I guess). Below are some zoomed in parts that show these trouble areas.
I did have to use the MRS instead of KMean clustering for the detection of the segments. I know you prefer KMeans due to performance, but that method always combined water and shadow in one segment near one of the dams. In that case they will also be detected as the same.
- Messages
- 1,588
- Country
Impressive! That reservoir is a good place to start as the water is fairly calm there being more of like a lake than a river. The tricky parts, as you note, will be "white water" due to rapids, waves or even sunlight reflection. On the other side is dark shadows on the water from over-hanging trees or tree shades on land. Yes, I do favor KMean mainly due to processing time and I could not see much of any difference in the results. If MRS provides good results that I would not have to go back and clean up, I could find times to run water detection if it takes way long... like run it if I am away at lunch or overnight. You get the idea. The last shot is remarkable - completely usable as is. How long did it take scenProc to process that one one?
Question just popped into my mind. Any advantage of having both MRS and Kmean steps in the same line of steps? Never thought of trying that. Or not possible or have any advantage.
It will be interesting to see results when you try the rivers on that same image.
Question just popped into my mind. Any advantage of having both MRS and Kmean steps in the same line of steps? Never thought of trying that. Or not possible or have any advantage.
It will be interesting to see results when you try the rivers on that same image.
- Messages
- 33,139
- Country
A run on the 1.3 GB image is around 40 minutes now with the MRS step.
I did another run with more sample points to catch the waves and the darker river. But that Also resulted in more false positives on land unfortunately.
I could remove most of the wave spots with a dilate/erode step, but that would also kill some of the jetties that are so nicely defected now.
I did another run with more sample points to catch the waves and the darker river. But that Also resulted in more false positives on land unfortunately.
I could remove most of the wave spots with a dilate/erode step, but that would also kill some of the jetties that are so nicely defected now.
- Messages
- 1,588
- Country
I wonder how MSFS does it?
- Messages
- 1,588
- Country
Yea, and the deep pockets ($$), to afford that. FYI - When I use KMEANs it takes me around 1 hour to process 25 of the images you are testing with. If MRS is 40 minutes, were looking at up to 17 hours to process. Now you can see why I went the Kmeans way. 17 hours would kill me... I would have to spend 2 days just on water. It's a "pickle" which way to go...
- Messages
- 33,139
- Country
When I did the vegetation detection for an entire state it often also took half a day at least to process (and I used the scenProcBatchRunner to process multiple areas in parallel). These steps just take time, but once you have the filter trained as you want you can just run it on an entire state (or at least a bigger area) and wait for the results. So as I see it, it does not matter if it takes a night or even a day or two in that case.
The choice between KMeans and MRS is a tradeoff between quality and processing time indeed. It's up to the developer to make that choice for himself.
A new development release with the changes I have made before will be online later this morning. I will send you my test TF2 file so that you can use that as a starting point for some experimentation yourself.
The choice between KMeans and MRS is a tradeoff between quality and processing time indeed. It's up to the developer to make that choice for himself.
A new development release with the changes I have made before will be online later this morning. I will send you my test TF2 file so that you can use that as a starting point for some experimentation yourself.
- Messages
- 1,588
- Country
Very interested to try the latest. I agree, if the results require little clean-up (hopefully none), a 1/2 day or overnight processing may well be worth it. I guess I will soon see. To be clear, the new process builds its learning from just the samples I provide for that session or will it retain previous sessions from other images and build that database? I see now reading your other post quite a few changes are taking place. I'll read-up and explore. After several weeks I will post my findings. There may be some additional questions along the way as I experiment. Thanks for taking the time to taking a further dive into this aspect of scenPROC.
Last edited:
- Messages
- 33,139
- Country
Hi,
I have not done anything with the idea of collecting it in a database yet. So you can see the TF2 file as your database for now, where you collect all samples. I am not sure if combining samples from different types of images would improve the results or not. Maybe I can try to do the WA area as well and then see if a combined TF2 with the sample points from CA and WA performs in a similar way or not.
I have not done anything with the idea of collecting it in a database yet. So you can see the TF2 file as your database for now, where you collect all samples. I am not sure if combining samples from different types of images would improve the results or not. Maybe I can try to do the WA area as well and then see if a combined TF2 with the sample points from CA and WA performs in a similar way or not.